物种分布模型的发展及评价方法

物种分布模型已被广泛地应用于以保护区规划、气候变化对物种分布的影响等为目的的研究。回顾了已经得到广泛应用的多种物种分布模型,总结了评价模型性能的方法。基于物种分布模型的发展和应用以及性能评价中尚存在的问题,本文认为:在物种分布模型中集成样本选择模块能够避免模型预测过程中的过度拟合及欠拟合,增加变量选择模块可评估和降低变量之间自相关性的影响,增加生物因子以及将物种对环境的适应性机制(及扩散行为特征)和潜在分布模型进行结合,是提高模型预测性能的可行方案;在模型性能的评价方面,采用赤池信息量可对模型的预测性能进行客观评价。相关建议可为物种分布建模提供参考。

The development and evaluation of species distribution models

Species distribution models (SDMs) have been used in various applications, such as conservation planning, determining the impact of climate change on species distribution, and others. SDMs allow construction of the correlation relationship between occurrence of a target species and environmental conditions (including bioclimatic and anthropogenic conditions). The correlation relationship can then be applied to the entire environmental space to predict the potential distribution of a target species. In the present study, we first review widely-used SDMs and summarize their evaluation approaches. Generally, SDMs can be classified into two categories according to the data required for construction of the correlation relationship, i.e., SDMs that predict the potential distribution of species based on presence-only records (PO models), and SDMs that use presence-absence records (PA models). If reliable absence records are available, PA models are suggested. Additional classification of SDMs is based on output format, namely, SDMs that give prediction results in the format of continuous probabilities (the higher the probability, the more suitable for distribution), and those with results in the format of binary values (1 for suitable and 0 for unsuitable). According to the various SDM output formats, SDM performance can be evaluated by threshold-independent (for models with continuous probabilities) and threshold-dependent (for models with binary prediction) strategies. Threshold-independent strategies can be realized by calculating values of maximum overall accuracy, maximum kappa, maximum vertical distance, area under the receiver operating characteristic (ROC) Curve (AUC), Gini index, point biserial correlation coefficient, mean square error, root mean square error, coefficient of determination, mean absolute prediction error, and others. Threshold-dependent strategies can be realized by calculating values of sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, negative likelihood ratio, true skill statistic, odds ratio, Yule''s Y, Yule''s Q, Phi coefficient, Kappa, normalized mutual information, extreme dependency score, and others. Based on existing problems related to the development, application and evaluation of SDMs, the present study suggests the following. (1) Because biased presence samples can influence the result of prediction, integrating a sample selection module in the SDM could improve the reliability of model prediction. (2) Bioclimatic variables (such as WorldClim) that are calculated from a Digital Elevation Model (DEM) may co-linearly correlate with each other, and such collinearity may result in overfitting when modeling the potential distribution of a species. As a result, selecting variables based on calculation of the Variance Inflation Factor (VIF) is a suitable means to avoid overfitting. (3) In addition to abiotic factors, biotic factors are also important determinants for species distribution. Thus, the use of biotic variables could improve the model results, although biotic factors are not easy to delineate within a geographic space. (4) The spatial and temporal extrapolation of SDMs, which deal with problems of species potential distribution at different geographic ranges and time points (past and/or future), respectively, are actually based on the assumption of an equilibrium relationship between the target species and environmental conditions. However, this assumption is challenged, because species have the abilities of adaption and dispersal. (5) The Partial AUC (PAUC) is suitable for evaluation of single model performance, and the Akaike Information Criterion (AIC) could provide an objective evaluation of the performance of several SDMs.